NEWS.md

NEWS

Changes in 3.2 (2026-10-30)

Feature removals and incompatible changes

• The JSON output of nvme show-regs, get-reg and get-property has one shape: the value of each register under its name, e.g. {"cc": 4587521}, with -v an object of the value and its decoded fields. The values are unsigned. Before, get-reg printed {"register: 0x14": [{"name": "cc", "value": ...}]}, get-property the value as a hex string, -v wrapped each register in a one element array, and a register with bit 31 set was negative.

• nvme show-regs, get-reg and get-property -v decode the fields of the registers as the specification names them, the same in the normal and the JSON output, which differed before. The composite fields CAP.CSS, CAP.AMS, CAP.CRMS and VS are split into their fields. The normal output of show-regs names the Version register vs, like the JSON output and the options, instead of version.

• nvme set-reg falls back to a Property Set only for a controller whose registers can't be mapped, a fabrics controller. Before, it set the property after any failed write. It refuses a register the host may not write and a value too large for a 32 bit register, and fails without a register to write instead of doing nothing.

• nvme help no longer lists the deprecated commands. They still work, with a warning.

• nvme-cli parses its command line itself instead of with getopt, so it behaves the same with every C library. Long options now need two dashes and their full name: single-dash long options such as -verbose and abbreviations such as --namespace for --namespace-id are no longer accepted. A single dash always starts short options, so -verbose is read as -v -e -r -b -o -s -e. Options and arguments can still be given in any order.

• Every command rejects an invalid --output-format value now, before it opens the device. Before, the commands which don't use the output format accepted any value.

• nvme <command> --help (also -h and -?) prints the help to stdout and exits with status 0. Before, it printed to stderr and exited with 1, like a usage error, which still does.

• Every command checks its operand: a command which takes a device fails without one, and an extra operand is an error. Before, extra operands were silently ignored, e.g. a device given to a command which takes none.

• --persistent force fails now. The value of an option which takes an optional value has to be attached to it: --persistent=force or -pforce. A separate value was never applied: --persistent force connected in mode auto and dropped force without a word, as the man pages of discover, connect-all and config create said. Now it is an extra argument, which is an error.

• Options which take an unsigned number reject negative numbers now. Before, -1 wrapped around to the largest value. The values which have a special meaning got names instead: every namespace id option accepts all for 0xffffffff and --timeout (also in nvme-cli.conf) accepts infinite. Both still accept -1. The help lists the names of an option. lm migration-recv requires --sel now, like the other lm commands.

• --sel of get-feature, the feat commands and the ocp and shannon commands which get a feature rejects numbers above 3 and accepts the names current, default, saved and supported. Before, it sent the reserved values 4 to 7 and truncated larger values. --output-format-version (also in nvme-cli.conf) rejects versions other than 1 and 2. Before, every version but 2 printed version 1.

• Numbers on the command line and in nvme-cli.conf are parsed strictly:

  • Text after the number is rejected. Before, the rest was ignored: --timeout=10s set a timeout of 10 milliseconds.
  • Leading whitespace, a + and a leading zero are rejected, also in the lists of numbers. Before, a leading zero made a number octal: --namespace-id=010 was namespace 8 and 08 was rejected. Write 10, or 0x10 for hexadecimal; 0 itself is fine.
  • A signed option rejects a value out of its range instead of wrapping around.
  • A floating point option takes a . as decimal point in any locale and rejects nan and inf.
  • Sizes with a suffix (4Ki, 1.5G) reject a value which overflows, take the digits after a . in decimal (1.09k was rejected as octal, 1.500k was 1000), and use a . in any locale as well.
  • An empty value no longer selects a value name.

• Errors and warnings which repeat a command line argument or a line of a config file write its control characters as \xNN, so it can't send escape sequences to the terminal.

• Global options mean the same for every command now: they are looked up before the options of a command, and no command uses their names any more.

  • The own --human-readable (-H) of fdp configs, lm migration-recv, shannon get-additional-feature and zns id-ns, and the own --verbose (-V) of sndk and wdc vs-internal-log, sed discover and zns report-zones are gone; the global --verbose (-v) does their job. -V is no longer accepted by these commands.
  • These options lost their short name, which a global option uses: config convert --output and rpmb read-data/write-data --address (-o), feat arbitration --hpw, feat temp-thresh --thsel, feat host-behavior-support --hdisns, lm set-cdq --hp, nbft show --hfi and samsung vs-internal-log --hide-progress (-H).
  • --human-readable and -H, the old spelling of --verbose, print a deprecation warning.

• A usage error prints the error, the synopsis and where to find the help instead of the whole help, which buried the error. A bad value names the option as it was given and the values it takes:

  fw commit: invalid value 'x' for -s: expected a number from 0 to 255
  Usage: nvme fw commit <device> [OPTIONS]
  Try 'nvme fw commit --help' for more information.
  

  The other errors of a command start with its name as well, e.g. id ns: /dev/nvme0: Permission denied, and a device which can't be opened prints no help any more. An unknown command and a bad global option point to nvme --help instead of printing the general help, which also followed a command that failed with ENOTTY. On a terminal, that name is red and the WARNING of a warning yellow. Set NO_COLOR to a non-empty value to turn colors off (https://no-color.org/).

• Options which take one of a few words check the word while parsing the command line, in any case and as a unique prefix like every value name, and the help and the completions list the words: show-topology --ranking, memblaze select-download --select, micron select-download --select, the --option of the micron vs-smbus-option, latency-tracking and vs-telemetry-controller-option, the micron --command of latency-tracking and latency-stats, micron vs-internal-log --type, ocp internal-log --telemetry-type, the --type of the sndk and wdc vs-internal-log and solidigm internal-log --type. An unknown word is a usage error which lists the words. ocp internal-log with an unknown telemetry type failed with status 0 before. samsung vs-internal-log --dump-type checks its list of dump types the same way, before it opens the device, and may be given more than once; the completions complete the type after the last comma.

• The help of connect, discover, connect-all, disconnect and config create lists the fabrics options they share under "Fabrics options", and the help of the feat commands --save and --sel under "Feature options".

• The command metadata (nvme utils dump-command-metadata) marks deprecated commands with the command which replaces them, deprecated options with the option which replaces them, and renamed options with their old name.

• The old option names --dhchap-secret, --dhchap-ctrl-secret and --fast_io_fail_tmo of the fabrics commands and --force of discover and connect-all print a deprecation warning which names the option to use instead.

• The deprecated tls-key command parses its options like every other command: it has a help, and abbreviated options such as --keyr are no longer accepted. It, gen-dhchap-key and check-dhchap-key reject extra operands now instead of dropping them.

• The human-readable stdout output of id-ctrl, id-ns, and most other identify/log-page/feature commands is now rendered through a computed-width table instead of hand-aligned printf() calls. Field names, values, and decoded bit breakdowns are unchanged, but column widths, spacing, and blank lines around them may differ from previous releases. stdout was never a stable, parseable interface; scripts that depend on its exact layout should switch to --output-format=json (-o json), which this change does not affect.

• The ScaleFlux plugin no longer supports the proprietary ScaleFlux kernel driver used with CSD 2000 drives. nvme-cli does not support proprietary kernel modules. The clean card feature is removed.

• nvme discover, nvme connect-all and nvme config create --discovery now default --persistent to auto instead of no. With this, the discovery controller stays connected after the log page fetch wherever the target reports EPCSD for that discovery log page entry. That lets the host automatically react to discovery AENs received on that discovery connection. Targets without EPCSD enabled remain unaffected. Opt out with --persistent=no. config create --discovery and converted discovery.conf lines now record persistent = auto. See nvme-discover(1) and nvme-config-create(1).

• nvme-discoverd is now built and installed by default, and it replaces the legacy udev/systemd NVMe-oF autoconnect. The daemon connects the controllers from the NBFT, from the fabrics configuration (nvme-fabrics.conf(5)) and from the Discovery Log Pages of the Discovery Controllers it connects to, including FC kickstart. It also keeps them connected and retries failed connections with backoff. For a Discovery Controller without an owner, for example one connected with nvme discover --persistent, it runs nvme connect-all on its events, as 70-nvmf-autoconnect.rules does. The new connections have no owner either. See nvme-discoverd(8).

  The nvme-discoverd and nvmf-autoconnect build options now both default to auto. nvme-discoverd is built when libsystemd v253 or later is available. The legacy autoconnect is built only when nvme-discoverd is not. To keep the old behavior, build with -Dnvme-discoverd=disabled.

  When nvme-discoverd is built, these legacy files are not installed: 70-nvmf-autoconnect.rules, nvmf-connect@.service, nvmf-connect.target, nvmf-autoconnect.service, nvmefc-boot-connections.service, nvmf-connect-nbft.service, the dracut config 70-nvmf-autoconnect.conf and the NetworkManager dispatcher script 80-nvmf-connect-nbft.sh. The key provisioning and vendor tuning udev rules are still installed.

  -Dnvmf-autoconnect=enabled installs both. The legacy udev rules are active as soon as they are installed. nvme-discoverd runs only when nvme-discoverd.service is enabled. While nvme-discoverd is active, nvmf-autoconnect.service and the FC kickstart skip their connect step.

• libnvme: the Identify Controller bit field constants for the OAES, BPCAP, MEC, OACS, FRMW, LPA and AVSCC fields now use the field names from the NVM Express Base Specification, Revision 2.4. The old names are removed. Code that uses them must be changed:

  • NVME_CTRL_OAES_*: FA is now FAN, ANA is ANACN, PLEA is PLEAN, LBAS is LSIAN, EGE is EGEAN, NS is NNSS, TTH is TTHR, ZD is ZDCN and DL is DLPCN. This also applies to the _SHIFT and _MASK values. The accessor macros NVME_CTRL_OAES_PLEALCN(), NVME_CTRL_OAES_LBASIAN(), NVME_CTRL_OAES_EGEALPCN() and NVME_CTRL_OAES_NNVMSS() are now NVME_CTRL_OAES_PLEAN(), NVME_CTRL_OAES_LSIAN(), NVME_CTRL_OAES_EGEAN() and NVME_CTRL_OAES_NNSS().
  • NVME_CTRL_BACAP_* is now NVME_CTRL_BPCAP_*.
  • NVME_CTRL_MEC_SMBUSME is now NVME_CTRL_MEC_TWPME. The spec renamed the bit to 2-Wire Port Management Endpoint.
  • NVME_CTRL_OACS_*: SECURITY is now SSRS, FORMAT is FNVMS, FW is FWDS, NS_MGMT is NMS, SELF_TEST is DSTS, DIRECTIVES is DIRS, NVME_MI is NSRS, VIRT_MGMT is VMS, DBBUF_CFG is DBCS, LBA_STATUS is GLSS, CMD_FEAT_LD is CFLS, HMLM is HMLMS and CTRL_SCOPED_CMD_FEAT_LD is CCFLS. The accessor macros NVME_CTRL_OACS_NMS_M() and NVME_CTRL_OACS_VMS_M() are now NVME_CTRL_OACS_NMS() and NVME_CTRL_OACS_VMS().
  • NVME_CTRL_FRMW_*: 1ST_RO is now FFSRO, NR_SLOTS is NOFS, FW_ACT_NO_RESET is FAWR and MP_UP_DETECTION is SMUD. NVME_CTRL_FRMW_NOFS also covers bit 3 of the Number Of Firmware Slots field, which NVME_CTRL_FRMW_NR_SLOTS missed.
  • NVME_CTRL_LPA_*: SMART_PER_NS is now SMARTS, CMD_EFFECTS is CSES, EXTENDED is LPEDS, TELEMETRY is TS, PERSETENT_EVENT is PES, LI0_LI5_LI12_LI13 is MLPS and DA4_TELEMETRY is DA4S.
  • NVME_CTRL_AVSCC_AVS is now NVME_CTRL_AVSCC_VSCF.

• The docs-build option is removed. -Ddocs always installs the docs committed in the tree. The new update-docs target regenerates them and copies them into the tree. Drop -Ddocs-build=true from the meson arguments: meson rejects an unknown option.

New: FreeBSD support

• nvme-cli and libnvme now build and run on FreeBSD. NVMe PCIe devices are supported through the nvme(4) driver; NVMe-oF fabrics, MI, nvme-discoverd, and nvme top are not yet available on this platform.

New: privilege separation and decoder sandbox

• Two technology previews reduce what a bug in the code which decodes the data of a device can do. Both are disabled by default. They are work in progress and may change, or be removed again if they turn out not to be useful or too complicated. See Documentation/PRIVSEP.md.

  • Privilege separation (-Dprivsep=enabled, Linux only): when root runs nvme-cli for a user, with sudo or with NVME_PRIVSEP_UID set, nvme-cli starts nvme-privsep-helper, which keeps root, and drops its own privileges to that user. The helper issues the NVMe commands. Root without SUDO_UID or NVME_PRIVSEP_UID keeps its privileges. Commands without a device, like the fabrics, keys and registry commands, and device commands which need root for other reasons still run as root.
  • Decoder sandbox (-Dsandbox=enabled, Linux and FreeBSD): the printers which decode device data run in a child process which can do little more than allocate memory and write to stdout and stderr. A crash in a decoder only ends the child. NVME_SANDBOX=0 turns it off, NVME_SANDBOX=1 turns it on in a build without the option.

New: nvme-keysd

• New daemon, nvme-keysd, a technology preview: it will put NVMe/TCP TLS pre-shared keys into the kernel's .nvme keyring before a connection needs them. Built with the new nvme-keysd meson option, disabled by default. It imports the PSKs of the entries with tls-key-source = systemd-creds from encrypted systemd credentials in /etc/nvme/creds at boot, and exits. It does not start when no entry has a key source. It needs systemd 257 or later. See nvme-keysd(8).

• New tls-key-source key in nvme-fabrics.conf. With tls-key-source = systemd-creds, tls-key is the name of a systemd credential, not the key. nvme connect-all and nvme-discoverd then connect with TLS and let the kernel find the PSK in the keyring. The default, inline, keeps today's behavior. See nvme-fabrics.conf(5).

nvme-cli

• nvme top -o json samples the I/O statistics over --delay seconds and prints the rates and the error, retry, failover, reset and reconnect counters of the subsystems, namespaces, paths and controllers as one JSON document, for scripts and monitoring tools. nvme top has a man page now, which describes the document.

• scripts/mcp/nvme-mcp is an MCP server on top of nvme top -o json. It lets an agent watch the multipath I/O of a host: it flags paths which are not optimized, are slow compared with their siblings, get an uneven share of the I/O, or whose error counters go up. Started with --allow-iopolicy, it also changes the I/O policy of a subsystem and compares the policies under a running workload. See scripts/mcp/README.md.

• nvme help <command> finds the command like nvme <command> does, also by a unique prefix, and nvme help <plugin> <command> shows the man page of a plugin's command.

• The shell completions offer a device only to commands which take one. The command metadata (nvme utils dump-command-metadata) reports the device operand of each command as schema version 2.

• The global options (--verbose, --output-format, --dry-run, ...) are accepted by every command. config validate, config status, config convert, intel lat-stats-tracking, memblaze lat-stats, memblaze smart-log-add-x, sndk vs-internal-log, wdc vs-internal-log and the sed commands did not accept them.

• The command metadata reports the type of the value of each option and the range of numbers an option takes. The shell completions use them: they complete paths for every option which names a file or directory, now also the --keyfile of the keys commands, --nbft-path and the logs of ocp internal-log, and offer the values of --sel. They no longer offer 0 to 3 for the --sel of the lm commands, which means something else there.

• The help and the shell completions of --output-format list the formats the command supports. Before, the completions offered tabular everywhere, and some commands listed formats they rejected.

• --dry-run now also skips nvme reset, nvme subsystem-reset, nvme ns-rescan, the nvme sed operations which change the drive, and the fabrics connects and disconnects. Before, only NVMe commands were skipped: these still reset the controller or subsystem, rescanned the namespaces, connected and disconnected controllers, and with dry-run set in nvme-cli.conf the sed operations still ran. nvme connect --dry-run -vv shows the connect arguments instead. nvme discover and nvme connect-all stop before connecting the Discovery Controller, as there is nothing to discover without a connection.

• nvme discover and nvme connect-all now send the keep-alive timeout to a persistent Discovery Controller. Before, the value was set too late. The DC got no keep-alive timeout, so it was never asked for a persistent connection or for AENs.

• nvme discover and nvme connect-all add the scope of a Discovery Controller connected through a scoped link-local IPv6 address, e.g. fe80::1%eth0, to the link-local addresses of its Discovery Log Page entries. Before, these had no scope, and on a host with more than one link the connect could use the wrong one. RDMA has no host_iface to select the link.

• New option nvme connect --discovery. It tells the kernel that the controller is a Discovery Controller, also when it has a unique NQN.

• nvme connect over TCP without --nr-io-queues uses one I/O queue per queue of the network interface, at most one per online CPU. The interface is --host-iface, or else the one of the route to the target. Its queues are the combined channels plus the smaller of the RX and TX channels, as ethtool -l shows them. With one queue, or if the count can't be read, the kernel default applies as before. nvme connect-all does the same for a connection of nvme-fabrics.conf without nr-io-queues.

• nvme sed revert now prints an extra revert LSP error when the kernel cannot revert the Locking SP. nvme sed password checks at run time, not at build time, whether the kernel can set the SID password, and skips that step if it cannot.

• A device, or a fabrics target, which returns log pages whose lengths and counts don't add up no longer makes nvme-cli read past its buffers, loop for minutes or leak memory. This covers id-ctrl and the FDP, host and AVE discovery, pull model DDC request, power measurement, reachability, dispersed namespace, LBA status, changed namespace list and persistent event logs. Found by fuzzing the device data.

• nvme log host-discovery, ave-discovery and pull-model-ddc-req read the records which follow the header now. Before, they printed zeros instead. host-discovery --all-host-entries reads the header of the right page as well.

• nvme log reachability-groups and reachability-associations decode every descriptor now. Before, only the first one was read from the right place. Their JSON output lists the IDs of a descriptor in an array, nsid and rgid. Before, all but the last ID were lost.

• nvme log persistent-event fails for a log larger than 4 GiB, which it can't read, instead of printing a part of it as the whole log.

• nvme sanitize-ns sends the Sanitize Namespace command to the namespace of the device, or to the one the new --namespace-id option gives. Before, it sent it with NSID 0, which the controller rejects.

• nvme set-feature sends the timestamp of --value little endian, and no longer in Command Dword 11. It refuses a value which doesn't fit: 32 bits, 48 for the timestamp. Before, it cut off the rest.

• nvme device-self-test --wait aborts the self-test on Ctrl-C again, and no longer waits for a self-test which is done before it reads the progress the first time. Without a self-test code, it fails when it can't read the progress, instead of reporting no running self-test.

• nvme sanitize --wait waits for the estimated time on big endian hosts too, and no longer fails on a sanitize without an estimated time, like exit-failure, which is done. --repeat=0 is refused.

• nvme format on a controller without --namespace-id tells how to format a namespace or all of them.

• nvme read, write and compare refuse a --data-size of more than 65536 logical blocks, which overflowed a buffer, and size the metadata for all blocks of --data-size without --block-count.

• nvme copy sends 0 as the expected reference tag of a range without --expected-ref-tags. Before, it was whatever was on the stack.

• nvme get-lba-status sends the namespace of the device without --namespace-id, instead of NSID 0, and refuses a --max-dw whose data length doesn't fit 32 bits.

• nvme feat num-queues --nsqr=N without --ncqr keeps the current number of completion queues now. Before, it requested a single completion queue and a wrong number of submission queues.

• nvme feat perf-characteristics no longer crashes after it sets the feature. It refuses an --attrl above 4064 bytes, which overflowed a buffer, and sends the attribute length little endian.

• nvme feat arbitration sets the feature only with the option of a field. Before, any option but --sel, like -o json, set it.

• nvme feat power-meas gets the feature with --uuid-index too.

• nvme feat timestamp clears the attributes of the timestamp it sets. Before, they were whatever was on the stack.

• nvme feat lba-range-type takes --guids as optional on every architecture, refuses a --num above 63, which set 64 zeroed ranges, and shows its own description in its help.

• nvme fw commit reports a commit which needs a reset to activate the image as a success. nvme fw download and nvme security recv send the --ish flag of an NVMe-MI command.

• nvme ns create --nsze-si and --ncap-si take the namespace granularity of a format with extended metadata, and the LBA format on big endian hosts, right, and align to a granularity above 4 GiB.

• The nvme zns zone management commands and zone-append work again: they sent their I/O commands to the admin queue. zone-mgmt-recv requires a --data-len which is a multiple of 4 for each action, and zone-append refuses more than 65536 logical blocks and no longer overflows the metadata buffer.

• nvme resv report reports a failed Identify Controller with its status, and fails.

• nvme io-mgmt recv requires a --data-len which is a multiple of 4, exits with success after writing the data to --data, and fails if it can't write all of it.

• nvme dir send and dir receive default to the namespace of a namespace device, and dir send refuses an --endir above 1.

• nvme get-reg --register=NAME[,NAME...] and nvme set-reg --register=NAME --value=VAL name the registers, with completion of the names, and --offset of get-reg, set-reg, get-property and set-property takes a name too, e.g. --offset=cc. The options of a register each of earlier versions, like get-reg --cc and set-reg --cc=VAL, still work, but are deprecated.

• nvme show-regs and get-reg on a fabrics controller show the properties in the right byte order on big endian hosts, and read the NSSD and CRTO properties too. show-regs leaves out the properties the controller doesn't implement, also in the JSON output. Before, it printed all registers of a PCIe controller, as -1 if they weren't properties. It fails for any other error of an optional property.

• nvme get-reg and get-property on a fabrics controller fail for a property the controller doesn't implement, and get-reg for a register which is no property. Before, they showed it as all ones, or failed only if the status had the DNR bit set. get-reg prints the status of a failed Property Get, and refuses an --offset which is no register, which read past a buffer from 0x1000 on.

• nvme show-regs --verbose prints the CRTO timeouts in milliseconds. Before, it dropped the half seconds.

Plugins

• The JSON output of nvme wdc vs-smart-add-log prints the counters, averages and percentages of the C1 performance statistics as 64 bit values. Before, a value above 2^31 - 1 came out wrapped.

• The wdc plugin no longer creates its dump directories with a broken mode that only root could write to.

• Fixed root command injection in WDC (vs-internal-log, drive-essentials, SN730 capture) and SanDisk (telemetry capture): a caller-controlled string reached a shell (system()), so a crafted input could run arbitrary commands as root.

• nvme seagate vs-smart-add-log prints the SMART logs of a drive once. Before, it printed those of legacy drives twice and those of other drives followed by logs which aren't theirs, and its JSON output could crash.

• The dera, ibm, memblaze, ocp, seagate, solidigm and wdc plugins handle unexpected values from the device and free the memory they used.

• WDC, SanDisk, Micron, Solidigm and Samsung plugins no longer spawn external tar, zip, cat, lsmod, uname or dmesg processes; they use libarchive or direct syscalls instead.

nvme-discoverd

• meson install no longer installs nvme-discoverd.conf, which overwrote the admin's settings. It installs a commented nvme-discoverd.conf.sample instead; copy it to nvme-discoverd.conf to change a setting. Without the file, every setting keeps its default.

• Discovery Controllers are now connected with --discovery and a keep-alive timeout of 30 s, or the configured keep-alive-tmo. No keep-alive timeout is used for persistent = no.

• epcsd-poll-interval-minutes, fc-kickstart-interval-minutes and dc-giveup-timeout move from [Global] to a new [Discovery] section in nvme-discoverd.conf. In [Global], they are now ignored with an "unknown key" warning. See nvme-discoverd(8).

• nvme-discoverd can discover Discovery Controllers through mDNS (TP8009), using systemd-resolved. It is off by default. Enable it with zeroconf = true in the [Discovery] section. The new mdns build option (auto by default) controls whether mDNS support is built. It requires libsystemd v258 or later.

• nvme-discoverd releases a controller it no longer wants, for example one removed from the configuration or from a DC's log page, or one the exclusion list now matches. It clears the registry owner and leaves the connection up. It no longer connects the IOCs of a DC that another orchestrator connected. After a restart, it reconnects the DCs it found through mDNS. A DC configured with a persistent setting now connects.

• nvme-discoverd also compares the host ID, not only the host NQN, when it matches a desired connection with an existing one.

• nvme-discoverd adds the IPv6 scope to link-local addresses. For a Discovery Log Page entry, the scope comes from the Discovery Controller's own address. For an mDNS result, it is the interface. RDMA needs this, because it has no host_iface to select the link.

• A connection unit of nvme-discoverd no longer disconnects another controller that reuses its device name. The unit records the sysfs inode of its device at connect time and disconnects only if the inode still matches. At startup, nvme-discoverd removes the state of controllers that are gone.

• At shutdown, a connection unit of nvme-discoverd disconnects only after the file systems mounted with _netdev are unmounted. Before, it could disconnect a controller while its file system was still mounted. A file system on an NVMe-oF namespace needs _netdev in fstab.

• nvme-discoverd resolves the host names in nvme-fabrics.conf in parallel, without blocking. Before, it resolved them one at a time, and a slow DNS server stopped the daemon. A failed lookup is tried again, first after 1 s, then less often, up to every 5 minutes. Before, the connection was skipped until the next reload. While the first lookups run, nvme-discoverd does not release any controller.

libnvme

• Built with liburing, libnvme sends a passthru command through io_uring only where the kernel takes it: admin commands on the char device of a controller, I/O commands on that of a namespace, and nothing on a block device. Before, every command on a namespace block device (/dev/nvme0n1) failed with "Invalid argument", and every admin command on a namespace char device (/dev/ng0n1) with "Inappropriate ioctl for device".

• A failed read of /dev/nvme-fabrics is no longer cached for the life of the global context, so a later connect retries it.

• New libnvmf_kernel_option_supported() and libnvmf_kernel_options_for_each() report the fabrics options the kernel lists in /dev/nvme-fabrics.

• The libnvme3.pc pkg-config file now reports the nvme-cli release version, e.g. 3.2. Before, it reported the library ABI version, 1.0.0 in every release. A project can now require a release with, for example, libnvme3 >= 3.2. The ABI version and the soname (libnvme3.so.1) do not change.

• A discovered controller gets the KX-HMAC-CHAP secrets (kxchap-secret, kxchap-ctrl-secret) of the [Subsystem] or [Discovery Controller] section in nvme-fabrics.conf with the same NQN, in the file of the Discovery Controller that reported it. This applies to nvme connect-all and nvme-discoverd. A section with no controller = line now has a use: it holds the secrets of a subsystem, wherever discovery finds it. Two sections with the same NQN in one file that set different secrets are rejected. New libnvmf_config_get_discovered_params() and libnvmf_context_set_config(), and config_discovered_params() in the Python bindings.

• The Copy source range entries carry the Expected Logical Block Tags (ELBT), Application Tag (ELBAT) and Application Tag Mask (ELBATM) little endian again, as in 2.x, except the 80 bit ELBT of formats 1h and 3h, which stays big endian. 3.0 and 3.1 sent them big endian, which nvme copy --expected-ref-tags and --expected-app-tags inherit. The byte order is not spelled out in the specification, see #4278.

Build

• New optional build dependency: libarchive (WDC, SanDisk, Micron, Samsung and Solidigm archiving), auto by default. Without it, the affected subcommands fail with a clear error instead of depending on whatever tar/zip happens to be on PATH. See BUILDING.md.

• The lm, sandisk, sed, wdc and zns plugins are now built on Windows. Without kernel Opal support, nvme sed only supports discover.

Changes in 3.1 (2026-09-18)

Feature removals and incompatible changes

• The minimum required Meson version is now 1.0.0, up from 0.62.0. Meson 1.0.0 was released in December 2022, so any reasonably current build environment should already meet it.

• nvme gen-hostnqn must now be run as root. The DMI and device tree files it reads are readable by root only. An unprivileged run returned a random identifier instead. See nvme-gen-hostnqn(1).

libnvme

• libnvmf_generate_hostid() now falls back to the local machine ID, after DMI and the device tree and before a random UUID. The machine ID itself is never published. The UUID is derived from it with HMAC-SHA256 and a fixed application identifier.

• A system UUID of all zeros or all ones is now discarded as invalid.

• A UUID-format host NQN read out of firmware-supplied NBFT data now has its UUID lower-cased on entry if the firmware got that wrong. All NQN comparisons remain plain, byte-exact, locale-independent string comparisons, per the Base Specification; only the value is normalized, at the one point it enters libnvme, not the comparison.

• NVMe-oF URI parsing now normalizes the nvme+<transport> scheme to lower case regardless of how it was spelled, matching the Boot Specification's requirement for that token. DHCP root-path and NBFT data are not consistent about its case.

• The API name NVME_CAP_NSSRC() changed to NVME_CAP_NSSRS(). Since the CAP property NVM Subsystem Reset Supported bit symbol name is described as NSSRS by the spec then the API name was not correct.

Build

• -Ddocs=man and -Ddocs=all install the man pages shipped in Documentation/ unless -Ddocs-build=true is also given. The section 5 and 8 pages were never written there, so those builds failed to configure. They are regenerated from now on, and a release aborts if any page is missing.

• Building the man pages now requires xmlto. It was optional, so a build with -Ddocs=man -Ddocs-build=true and no xmlto installed succeeded and produced no man pages at all.

Changes in 3.0 (2026-09-07)

Feature removals and incompatible changes

• A new INI-format connection configuration parser and writer have been added (nvme-fabrics.conf + nvme-fabrics.conf.d/ drop-ins, replacing config.json/discovery.conf). See libnvme/design/CONFIG.md for the format. Existing configuration can be explicietly converted by 'nvme config convert' or implicitly by the first call of any fabric command.

• Key management has moved into a new nvme keys plugin. gen-dhchap-key, check-dhchap-key, gen-tls-key, check-tls-key, and tls-key are now nvme keys gen-kxchap-secret, check-kxchap-secret, gen-tls-psk, check-tls-psk, insert-tls-psk, import, export, and revoke. The old commands still work as deprecated aliases, except check-tls-key --insert, which has no equivalent there. Use nvme keys insert-tls-psk instead. The DH-HMAC-CHAP naming is also renamed to KX-HMAC-CHAP throughout, matching TP4201. See nvme-keys-gen-kxchap-secret(1) and nvme-keys-check-kxchap-secret(1).

• NVMe Base Specification 2.4 (ECN122) renames the Reservation Report data structures from "Registered Controller" to "Registrant" throughout. struct nvme_registered_ctrl and struct nvme_registered_ctrl_ext are now struct nvme_registrant and struct nvme_registrant_ext; the regctl/regctl_ds/ regctl_eds fields of struct nvme_resv_status are now regstrnt/registrant_ds/registrant_eds. nvme resv-report output changes to match: the regctl/regctls/regctlext labels and JSON keys are now regstrnt/registrants/registrantext. Update any code or scripts that reference the old names.

• NVMe Base Specification 2.4 (TP4150) renames the "Namespace Attribute Changed" asynchronous event to "Attached Namespace Attribute Changed", the "Changed Namespace List" log page to "Changed Attached Namespace List", and the "Namespace Attribute Notices" notice to "Attached Namespace Attribute Notices" (to distinguish them from the new Allocated Namespace variants added alongside them). NVME_AER_NOTICE_NS_CHANGED is now NVME_AER_NOTICE_ATTACHED_NS_CHANGED, NVME_LOG_LID_CHANGED_NS is now NVME_LOG_LID_CHANGED_ATTACHED_NS, and NVME_CTRL_OAES_NA* is now NVME_CTRL_OAES_NSAN* (matching the spec's new OAES mnemonic). nvme discover --log-id=changed-ns still works as a deprecated alias for the new changed-attached-ns.

• NVMe Base Specification 2.4 (TP4193) renames the "Create Exported NVM Subsystem" and "Manage Exported NVM Subsystem" admin commands to "Manage Exported NVM Subsystem Receive" and "...Send" respectively. nvme_admin_create_export_nvms is now nvme_admin_manage_export_nvms_receive, and nvme_admin_manage_export_nvms is now nvme_admin_manage_export_nvms_send. The opcode values (0x2a, 0x2d) are unchanged.

• NVMe Management Interface Specification 2.2 (TP6039) adds a Controller Health Status Changed Flags field to the Controller Health Data Structure, and renames the NVM Subsystem Health Data Structure's Composite Controller Status field to Composite Controller Status Flags, to distinguish per-bit change flags from the status/state bits they report. In libnvme, struct nvme_mi_ctrl_health_status gains a chscf field (and enum nvme_mi_chscf) carved out of what was previously trailing reserved space, so its size is unchanged; and struct nvme_mi_nvm_ss_health_status's ccs field/enum nvme_mi_ccs are renamed to ccsf/enum nvme_mi_ccsf with bit names getting an F suffix (e.g. NVME_MI_CCS_RDY is now NVME_MI_CCSF_RDYF) and a new TCIDAF bit.

• nvme gen-dhchap-key emits the secret it was given rather than the key transformed from it, so its output differs whenever --hmac is non-zero. --nqn/-n fed only that transform and is now accepted with a warning and ignored. See nvme-gen-dhchap-key(1).

• nvme gen-dhchap-key rejects a --secret with an odd number of hexadecimal characters. nvme-cli 2.x padded the trailing one with a zero and emitted a secret that was never given: 63 characters ending ...bababa produced ...ababab0a.

• nvme check-tls-key used to print the TLS PSK identity as the only line on stdout; it now prints Configured PSK is valid (HMAC <id>, length <len>) first, the identity second, and whether that identity is already in the keyring third, so a script reading the identity with head -1 gets the wrong line. Its --insert/--keyfile are gone; use nvme keys insert-tls-psk. See nvme-check-tls-key(1).

• nvme gen-tls-key rejects a --secret with an odd number of hexadecimal characters, reports Inserted TLS PSK <serial> where 2.x said Inserted TLS key <serial>, and prints its --secret diagnostics -- including Skipping excess secret bytes -- to stdout ahead of the key rather than to stderr. See nvme-gen-tls-key(1).

• nvme check-dhchap-key reports on the secret rather than on the key: Key is valid, Key is loaded and Key is not loaded are now Secret is ..., and the validation errors say secret where they said key. Exit codes are unchanged, but a script matching the old wording needs updating. See nvme-check-dhchap-key(1).

• All nvme *-log commands (smart-log, ana-log, error-log, fw-log, telemetry-log, self-test-log, sanitize-log, and the rest of the get-log-page family, get-log excepted) have moved into a new nvme log plugin, dropping the redundant -log suffix, e.g. nvme smart-log is now nvme log smart. The old top-level commands still work as deprecated aliases.

• All nvme id-*/nvm-id-*/list-* Identify commands (id-ctrl, id-ns, id-ns-granularity, id-ns-lba-format, list-ns, list-ctrl, nvm-id-ctrl, nvm-id-ns, nvm-id-ns-lba-format, primary-ctrl-caps, list-secondary, cmdset-ind-id-ns, ns-descs, id-nvmset, id-uuid, id-iocs, id-domain, and list-endgrp) have moved into a new nvme id plugin, e.g. nvme id-ctrl is now nvme id ctrl. The old top-level commands still work as deprecated aliases.

• Namespace management commands (create-ns, delete-ns, attach-ns, detach-ns, get-ns-id) have moved into a new nvme ns plugin, e.g. nvme create-ns is now nvme ns create. The old top-level commands still work as deprecated aliases.

• nvme ns create, nvme ns delete, nvme ns attach, and nvme ns detach (and their deprecated create-ns/delete-ns/ attach-ns/detach-ns aliases) no longer print a confirmation message on success by default; nvme ns create also no longer prints the assigned nsid. Both are now gated behind -v/--verbose. The exit code alone indicates success/failure; the assigned nsid can still be found with nvme ns list or nvme id ctrl. See nvme-ns-create(1).

• Reservation commands (resv-acquire, resv-register, resv-release, resv-report) have moved into a new nvme resv plugin, e.g. nvme resv-acquire is now nvme resv acquire. The old top-level commands still work as deprecated aliases.

• NVMe-MI commands (nvme-mi-recv, nvme-mi-send) have moved into a new nvme nvme-mi plugin, e.g. nvme nvme-mi-recv is now nvme nvme-mi recv. The old top-level commands still work as deprecated aliases.

• I/O Management commands (io-mgmt-recv, io-mgmt-send) have moved into a new nvme io-mgmt plugin, e.g. nvme io-mgmt-recv is now nvme io-mgmt recv. The old top-level commands still work as deprecated aliases.

• Directive commands (dir-receive, dir-send) have moved into a new nvme dir plugin, e.g. nvme dir-receive is now nvme dir receive. The old top-level commands still work as deprecated aliases.

• Security Send/Receive commands (security-send, security-recv) have moved into a new nvme security plugin, e.g. nvme security-send is now nvme security send. The old top-level commands still work as deprecated aliases.

• Firmware commands (fw-commit, fw-download) have moved into a new nvme fw plugin, e.g. nvme fw-download is now nvme fw download. The old top-level commands still work as deprecated aliases.

• The RPMB command has moved into a new nvme rpmb plugin and, since RPMB is not part of the NVMe base specification, has been split into one subcommand per action instead of a --cmd=<action> flag: nvme rpmb --cmd=info is now nvme rpmb info, and likewise program-key, read-counter, read-data, write-data, read-config, and write-config. Unlike the other command moves above, the old nvme rpmb --cmd=... form no longer works at all; there is no deprecated alias. See nvme-rpmb-info(1) and friends.

• Deprecated commands (currently the old nvme keys, nvme log, nvme id, nvme ns, nvme resv, nvme nvme-mi, nvme io-mgmt, nvme dir, nvme security, and nvme fw aliases above) are enabled by default; they can be turned off at build time with -Ddeprecated-cmds=disabled. nvme help now lists them in a separate "deprecated" section instead of alongside current commands, and running one prints a warning that it will be removed in the next major version.

• The project now builds with -std=gnu11 instead of -std=gnu99, raising the minimum required compiler support from C99 to C11. A toolchain new enough for C11 is now required to build nvme-cli and libnvme.

• nvme disconnect-all with no options no longer disconnects every fabric controller. It now only disconnects controllers with no recorded owner in the new ownership registry. A controller owned by another orchestrator is silently skipped. To restore the old disconnect-everything behavior.

• libnvme itself no longer resolves hostnames. libnvmf_add_ctrl() and libnvmf_connect_ctrl() now fail immediately on a hostname traddr/host_traddr instead of resolving it internally. The caller is responsible for resolving first.

• --dump-config has been removed from nvme discover, connect-all, and connect.

• In libnvme, the per-command nvme_<cmd>() wrapper functions and their struct nvme_<cmd>_args argument bundles have been replaced by an nvme_init_<cmd>() helper that fills a caller-owned struct libnvme_passthru_cmd, submitted separately with libnvme_exec_admin_passthru() / libnvme_exec_io_passthru() (or the async equivalents).

• The public NBFT parsing API has moved from the libnvme_* namespace to libnvmf_*, and every function now takes the global context as its first argument (for example, libnvmf_read_nbft(ctx, ...)). Update both the prefix and the call signature.

• The -Dpdc-enabled meson build option has been removed. Whether a discovery controller connection is kept persistent is now decided at runtime, via --persistent/nvme-fabrics.conf, not at build time.

New: Windows support

• nvme-cli and libnvme now build and run on Windows (MSYS2 UCRT64). Direct PCIe access via ioctl works; NVMe-oF fabrics, MI, nvme-discoverd, and nvme top are not yet available on this platform.

New: nvme top

• nvme top is a new interactive, top-like dashboard for NVMe devices. It lists subsystems and lets you drill into one to see live namespace, path, and controller stats, including command retry/error counts and multipath failover/reconnect counts.

New: ownership registry and exclusion list

• A new ownership registry (nvme registry, backed by /run/nvme/registry/) records which orchestrator owns each connected controller. This is what lets disconnect-all (above) and future orchestrator tooling avoid tearing down a connection another component depends on. See libnvme/design/REGISTRY.md.

• A new system-wide exclusion list (nvme exclusion, backed by /etc/nvme/exclusions.conf and exclusions.conf.d/ drop-ins) lets an administrator block specific controls from being auto-connected. It's aimed primarily at auto-discovered controllers, which have no config entry to remove in order to suppress an unwanted connection. See libnvme/design/EXCLUSIONS.md.

nvme-cli

• A new global config file, /etc/nvme/nvme-cli.conf, sets machine-wide defaults for global options like --timeout, --output-format, --no-retries, and --no-ioctl-probing. See nvme-cli.conf(5).

• nvme disconnect now accepts the same identifying options as nvme connect (--nqn, --transport, --traddr, --host-iface, --hostnqn, --hostid). A new -x/--exclude also writes a matching entry to the exclusion list before disconnecting. See nvme-disconnect(1).

• nvme utils dump-command-metadata prints the full command and option tree as JSON. It is meant to drive shell-completion generation.

• nvme connect gained --idempotent and --devid-file. --idempotent makes connecting to an already-connected controller succeed instead of erroring. --devid-file writes the resulting nvmeX device name to the given file on success, so a caller that doesn't know the device name at connect time (for example, a systemd unit spawned before the device exists) can look it up afterward instead of scraping dmesg or polling sysfs.

• nvme config-convert converts the legacy config.json and discovery.conf files to the new INI-format nvme-fabrics.conf. By default it reads the system paths and writes to /etc/nvme/nvme-fabrics.conf. --config overrides the config.json path, --output the destination, and --force allows overwriting an existing target. On success each converted legacy file is renamed to <name>.converted so running the command again is a safe no-op. See nvme-config-convert(1).

libnvme

• The library has been renamed from libnvme to libnvme3, and the previously separate libnvme-mi has been merged into it. The shared library SONAME is libnvme3.so.1. Headers install under include/libnvme3/ and man pages under a versioned path, enabling parallel installation alongside libnvme v1 packages. The Python binding is now named libnvme3. Packages, build systems, and Python scripts that reference the library, its headers, or the Python module by name must update.

• nvme-fabrics.conf entries can now record a persistent setting per discovery controller, matching the --persistent CLI option.

• New diagnostic accessors report per-path, per-namespace, and per-controller command retry/error counts, multipath failover count, and controller reset/reconnect counts. These are always read live rather than cached, and back nvme top.

• The Python bindings' object-oriented surface has been polished: connected() and is_registration_supported() are now read-only properties (connected, registration_supported), registration_ctrl() is renamed registration_control(), and set_symname() has been removed. Set host.hostsymname directly instead.

• The NBFT parser has been updated for NVMe Boot Specification rev. 1.3, including Security Profile Descriptors, and hardened against malformed or malicious NBFT ACPI tables.

• ioctl version probing is now deferred until the first command on a transport handle, instead of happening eagerly on open. The probe prefers the 64-bit ioctl and falls back to the 32-bit one only if it isn't supported; the result is cached per handle. The new --no-ioctl-probing global option skips probing and forces the 32-bit ioctl.

• Controller and namespace sysfs attributes are now read lazily, on first access, instead of all at once when the topology tree is scanned. This cuts overhead when scanning large fabric setups.

• Connect and discover now share a common struct libnvmf_context internally, replacing the old discovery-only context. Relevant to developers extending the fabrics API, not to CLI end users.

• A new internal shared/ static library holds code with no natural home in nvme-cli, libnvme, nvme-discoverd, or the Python bindings (INI parsing, base64, CRC32, filesystem/network helpers).

• Read-only Python bindings for the new config format (config_read(), config_validate()) are available now for early adopters (e.g. nvme-stas) that want to start reading the new format ahead of the CLI's own switch-over.

• The accessor generator that produces libnvme's getter/setter boilerplate (accessors.c/.h) has been rewritten in Python, with the struct annotations it reads now living in-source in private.h instead of a separate spec file. Relevant to developers extending libnvme's public structs, not to CLI end users.

• <nvme/nvme-cmds.h> and <nvme/nvme-types.h> have each been split into per-spec-section sub-headers (nvme-cmds-base.h, nvme-cmds-fabrics.h, nvme-cmds-mi.h, similarly for nvme-types). The top-level headers still exist and include all sub-headers, so code that includes them is unaffected. Code that already included a sub-header by its old combined-file path must update the include path.

• nvme_root_t has been replaced by struct libnvme_global_ctx * throughout the public API. Code that stored or passed nvme_root_t must update to the new type.

• A new transport handle abstraction (struct nvme_transport_handle, libnvme_transport_handle_*) decouples command issuing from the underlying transport. A handle may wrap either a direct ioctl file descriptor or an MI endpoint.

• An explicit async passthru API has been added. libnvme_submit_admin_passthru() / libnvme_submit_io_passthru() submit a command without waiting; libnvme_wait_passthru() / libnvme_reap_passthru() collect the result later. The synchronous path is libnvme_exec_admin_passthru() / libnvme_exec_io_passthru().

• libnvmf_host_get_ids() has been promoted from an internal helper to a public API. It resolves the hostnqn and hostid from all available sources and is the recommended way for callers to obtain host identity before connect or discover.

• The implicit hostnqn/hostid lookup inside the fabrics connect and discover code paths has been removed. Callers must now resolve host identity before connecting; use the newly public libnvmf_host_get_ids() helper or set values explicitly with libnvme_global_ctx_set_hostnqn() / libnvme_global_ctx_set_hostid().

• The sizeof_args backward-compatibility macro and the old struct-based command-argument shim have been removed. The macro returned incorrect values on 32-bit architectures; there is no intention to fix it.

• The platform-specific filter helpers (nvme_filter_*) are no longer exported from libnvme. Only the libnvme_scan_* interfaces remain public. Plugins that called filter helpers directly must be updated to use the scan API.

• libnvme_ctrl_match_config() has been removed. It was exported but had no callers outside the library.

• libnvmf_ctrl_get_fabrics_config() has been removed. The individual field accessors generated by the new accessor machinery provide equivalent access without exposing the internal nested struct.

• nvme_mi_ctrl_t has been removed. MI functionality is now accessed through struct nvme_transport_handle, the same abstraction used for direct ioctl access; the transport is chosen at handle construction. nvme_mi_admin_identify() and nvme_mi_admin_identify_partial() are gone; use libnvme_ctrl_identify() with a transport handle opened for MI.

• Several deprecated identifiers have been removed for 3.0: the duplicate NVME_SC_FEAT_IOCS_COMBINATION_REJECTED define, the old notification-mask enum names, MI backward-compatibility #defines, and nvme_cmd_get_log_telemetry_host_lsp (renamed to nvme_log_telemetry_host_lsp). Update any code using these names.

• The environment-variable configuration knobs for libnvme have been removed. LIBNVME_MI_PROBE_ENABLED, LIBNVME_TEST_BASE_DIR, and similar variables no longer have any effect. Replace them with explicit setter calls on the global context: libnvme_set_probe_enabled(), libnvme_set_test_base_dir(), libnvme_set_test_sysfs_dir(), libnvme_set_force_4k(), etc.